Jyothika Pratheep

Posted on Oct 12

MonsoonMap: Walk Your Streets. Expose the Problems. Make the Evidence Count.

#devchallenge #hf26challenge

What I Built

MonsoonMap is a civic-tech field companion that turns outdoor neighborhood walks into structured, human-verified environmental evidence.

Every monsoon or heavy rain shower transforms our streets into obstacle courses:

Pedestrian curb ramps submerge under four inches of stagnant water, forcing wheelchair users and parents with strollers directly into active vehicular traffic.
Roadside stormwater grates choke on plastic wrappers and leaf silt, causing runoff backflow.
Concrete pavers dislodge into muddy trip hazards.

Yet, most civic frustration dissolves into vague social media complaints ("my street is flooded"). Municipal ward engineers and community assemblies lack structured, verifiable physical evidence to prioritize interventions.

How it gets people off the screen and into the world

MonsoonMap was intentionally designed against continuous screen time:

Targeted Outdoor Walk Sessions: Citizens start a 10, 20, or 30-minute walk focused on a specific environmental issue (Rainfall & Waterlogging, Drainage & Inlets, Footpath & Pavement, or Runoff Litter).
Dedicated "Pocket Mode": The moment your walk begins, you tap Pocket Mode. The screen dims into a minimalist, low-power heads-up display. The app tells you: "Keep your phone in your pocket and observe your surroundings." MonsoonMap keeps track of your time in the background so you can look at the real world.
Contextual Observation Checklists: Instead of distracting forms, small field prompts guide your eyes: "Look for standing water forcing pedestrians off sidewalks", "Check if debris is occluding >50% of the intake grate".
Fast Capture & Pocket: When you discover an obstacle, you take a quick photograph, add a single-sentence note, and put your phone right back into your pocket.
Post-Walk Human Review: Only after you finish your walk do you review the findings, correct any AI classifications, and generate an evidence document.

Who is it for?

Everyday walking residents, neighborhood associations, ward committees, pedestrian safety advocates, and local civic volunteer groups who want evidence that counts.

Demo

Deployed URL: https://monsoonmap.vercel.app/
Full Source Code: https://github.com/jyothikacodes/monsoonmap

The Core Walkthrough

Outdoor Walk Setup: Choose a 20-minute route with a focus on Rainfall & Waterlogging.
Pocket Mode & Street Walking: Put the phone in your pocket; the timer runs unobtrusively in the background.
Field Capture: Spot a submerged curb ramp, snap a picture, and jot: "Puddle 4 inches deep covering entire crosswalk ramp, forcing pedestrians with strollers into traffic."
Client-Side Compression: High-res smartphone camera photos (8–15 MB) are compressed on-device in HTML5 Canvas to ~180 KB in milliseconds.
Open-Weight AI Structuring: The local vision model extracts observable physical factors, isolates facts from assumptions, and suggests follow-up checks.
Citizen Verification: You review the structured findings, adjust severity, confirm your field note, and approve the observation.
Map & Civic Evidence Report: View the geotagged record with custom color-coded category markers on OpenStreetMap, check for nearby clusters within 150m, and export a clean printable PDF or copyable Markdown report for your local ward assembly.

Code

The complete source code is open source under the MIT License:

jyothikacodes / monsoonmap

Walk your streets. Expose the problems. Make the evidence count. Hacktoberfest x DEV Week 1 Touch Grass Challenge.

MonsoonMap

Walk your streets. Expose the problems. Make the evidence count.
Built for the Hacktoberfest x DEV Week 1 “Touch Grass” Challenge.

🌐 Live Deployed Web Application: https://monsoonmap.vercel.app

MonsoonMap encourages people to spend time outdoors observing their surroundings and documenting environmental and infrastructure problems—particularly after rainfall.

During a short outdoor observation walk, citizens document waterlogging, choked stormwater drains, accumulated runoff litter, and damaged footpaths. MonsoonMap uses an open-weight vision-language model (Gemma 3 Vision via Ollama) to turn raw field observations into structured, human-reviewable civic evidence reports.

Designed from the ground up for pocket carry: the interface encourages users to put their phones away between observations, never rewarding continuous screen time.

🏛️ Why MonsoonMap? The "Touch Grass" Civic Mission

When heavy rains hit a neighborhood, ordinary sidewalks transform into obstacle courses. Submerged curb ramps force wheelchair users and parents with strollers into vehicular traffic lanes; plastic litter…

View on GitHub

Key Architectural Components

src/lib/ai-analyzer.ts: Local Ollama adapter connecting to Gemma 3 Vision with strict JSON schema validation and transparent fallback engineering.
src/app/walk/page.tsx: Outdoor walk interface featuring duration goals, focus checklists, and battery-friendly Pocket Mode.
src/components/ObservationsMap.tsx: Interactive Leaflet OpenStreetMap integration with custom SVG markers and cluster detection.
src/lib/db.ts: Embedded SQLite persistence powered by Node.js built-in node:sqlite.
src/lib/client-utils.ts: In-browser canvas image compression, Haversine proximity calculations, and markdown evidence report generator.

How I Built It

MonsoonMap is built with Next.js 16 (App Router), React 19, Tailwind CSS v4, Leaflet / OpenStreetMap, and Node.js node:sqlite.

The Open-Source AI Pipeline: Gemma 3 Vision

At the heart of MonsoonMap is Google's open-weight Gemma 3 Vision model (gemma3:4b-instruct) running locally through Ollama.

Disciplined Evidence Separation

Civic reports require rigorous physical discipline. We engineered our prompts and schema to strictly separate directly observable physical facts from speculative assumptions:

Observable Physical Indicators: Identifies visible factors such as "Standing water covers 100% of the curb ramp cut" or "Intake grate surface is >70% occluded by plastic and silt".
Refusal to Speculate: If the image shows standing water, the system identifies the puddle but explicitly refuses to claim that the underground municipal drainage pipe is broken unless visual evidence shows active backflow.
Verification Questions for the Human: Prompts the walking citizen with clarifying questions: "Does the water persist after 4 hours of dry weather?", "Is the blockage superficial leaf debris or an internal culvert choke?".

{
  "suggestedCategory": "waterlogging",
  "evidenceSummary": "Surface water accumulation visible over pedestrian footpath / curb corridor.",
  "observableSeverityIndicators": [
    "Water pool forces pedestrians into vehicular traffic lane",
    "Estimated standing water extends across entire walkable path"
  ],
  "severityLevel": "high",
  "confidenceScore": 0.88,
  "confidenceRationale": "High confidence based on visible path blockage and vehicle lane proximity.",
  "suggestedFollowups": [
    "Check whether ponding dissipates within 3-4 hours after rain ceases",
    "Inspect nearby curb drop-inlets for surface occlusions"
  ],
  "questionsForHumanVerifier": [
    "Is the standing water deeper than ankle height at its lowest point?"
  ]
}

Transparent Fallback Engineering

If a user runs MonsoonMap on hardware without Ollama or GPU resources, the application never pretends AI is working.

When Ollama is running with Gemma 3: Displays Gemma 3 Vision (Live Inference).
When offline: Falls back to deterministic heuristic parsing with an explicit badge: Gemma 3 Vision (Offline Demo Fallback). The entire civic workflow—walk timer, checklists, compression, human review, mapping, proximity detection, and PDF evidence export—remains 100% functional offline.

Why Does Open Innovation Matter?

When documenting local public infrastructure, using proprietary hosted AI APIs presents serious ethical and practical hurdles:

Civic Privacy & Non-Surveillance: When residents document their daily walking routes and neighborhood corners, those photographs should never be sent to commercial cloud providers for training or telemetry. Open-weight models running on local hardware ensure that community observations remain strictly under the user's control.
Zero Vendor Lock-In & Perpetual Access: Community assemblies and civic volunteer groups operate with zero software budget. They cannot depend on proprietary APIs that require paid subscriptions or suddenly deprecate endpoints. Open-weight models like Gemma 3 guarantee permanent availability.
Auditable Civic Trust: When presenting field evidence to municipal authorities, the structuring pipeline must be transparent. Open models allow full scrutiny of prompt constraints, eliminating commercial black-box bias.

Prize Categories

Touch Grass (Week 1): Primary category — outdoor walk mode, contextual checklists, screen-dimming Pocket Mode designed to put phones away, and post-walk evidence verification.
Open-Weight Models: Local Gemma 3 Vision integration via Ollama with strict evidence schemas and transparent diagnostic fallback.

DEV Community